PII-Tracer

PII-Tracer

Finds personal data in chat conversations and masks it, using a 600M-parameter encoder with a token tagger and a sensitivity score. It suits teams that clean conversation logs before use.

Made by perplexity-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: PII-Tracer is a model that spots personal data in conversations so it can be masked. It comes from the PII-TRACE paper on context-aware detection in multi-turn chats. Under the hood is a bidirectional Qwen3 encoder of about 600M parameters with two outputs. One tags personal data in the text, and the other scores how sensitive the whole conversation is.

Who it suits: It suits teams that handle chat logs or LLM conversations and need to remove names, contact details and secrets before storing, sharing or training on them. You need some Python to run it.

Industry | Who uses it | How they use it | Example
AI product teams | Data and ML engineers | Mask personal data in conversation logs before training or review | Cleaning stored chat transcripts
Customer support | Support operations | Redact customer details from conversations | Hiding emails and phone numbers in a ticket thread
Compliance | Privacy officers | Flag sensitive conversations for a closer look | Ranking chats by their sensitivity score
Research | Privacy researchers | Test context-aware detection on multi-turn text | Reading the model's tags on sample chats

People who work with long documents can still use it, but must split texts first, because inputs are cut at 4096 tokens.

Where to use it:

• Python with Transformers 5.2 or newer, on Windows, macOS or Linux
• vLLM, through the separate PII-Tracer-vLLM packaging
• llama-server, through the GGUF packaging
• MLX, through the MLX packaging, which needs your own loader

Use case | Who | What they get back
Finding personal data | Data engineers | Spans with a label, a start and an end position
Masking text | Support and compliance teams | The text with each span replaced by a tag
Sensitivity scoring | Privacy officers | One sensitivity value for the whole conversation

How to set it up:

• Python on your computer
• PyTorch and Transformers 5.2 or newer
• Permission to load the model with trust_remote_code

1. Install the Hugging Face hub package.
2. Download the model files.
3. Install torch and a recent Transformers, then load the model in Python.

pip install -U huggingface_hub

hf download perplexity-ai/PII-Tracer

pip install torch "transformers>=5.2"

How to use it: Load the model with trust_remote_code, then call predict on a text. It returns the detected spans and a sensitivity value. Call mask to get the text with personal data replaced by tags. Split texts longer than 4096 tokens before calling predict.

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "perplexity-ai/PII-Tracer", trust_remote_code=True
)

text = ("Hi, I'm Daniel Whitfield, you can reach me at "
        "daniels@meridiancap.com or 415-555-0123.")
spans, sensitivity = model.predict(text)
for s in spans:
    print(s.label, (s.start, s.end), text[s.start:s.end])

print(model.mask(text))

Tool | What it does
predict | Returns the personal data spans and a sensitivity value
mask | Returns the text with each span replaced by a tag
save_pretrained | Saves the model and tokenizer so the folder loads on its own

Example: You pass a short chat message that includes a name, an email address and a phone number. predict lists three spans, each with a label such as private_person, private_email or private_phone. mask then returns the message with those parts replaced by tags in capital letters.

For newcomers and experts: Masking replaces private details with a neutral tag so the rest of the text stays usable. A good first step is to run the card's sample on one made-up message and compare the output.

The token head tags text with BIOES labels over 9 categories, decoded by a constrained Viterbi step. The categories are person, account number, URL, date, address, email, phone, secret and other personal data. The sensitivity head reads the mean of the hidden states for a conversation-level score.

Conversation text goes in
Bidirectional Qwen3 encoder reads it
Token head tags spans, Viterbi decodes
Sensitivity head scores the chat

Good to know:

• License: MIT.
• Cost to run: none named. The steps need no API key or paid service.
• Data it reaches: it downloads files from the Hugging Face Hub, then processes the text you give it. Loading uses trust_remote_code, so it runs code from the repository.
• Not for: inputs beyond 4096 tokens without splitting them first.

On Hugging Face, 13 November 2026

• Task: token classification
• Parameters: 596M
• Weights: about 1.2 GB in BF16; about 298 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 8,506
• Likes: 97
• License: mit
• Languages: en
• Quantized versions: 5
• Fine-tunes: 3
• Tags: safetensors, pii_masking, token-classification, privacy, qwen3, multilingual

Links

• Model: https://huggingface.co/perplexity-ai/PII-Tracer
• Files: https://huggingface.co/perplexity-ai/PII-Tracer/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download perplexity-ai/PII-Tracer

Then follow the usage in its model card. Nothing in this kit runs by itself.